AMD CTO Phil Hester filled in some of the mind-expanding particulars of AMD’s server strategy during sessions at the opening of AMD’s Fab 39 in Dresden, Germany. Hester came into AMD with ambitious goals, the most ambitious of which is to carry AMD’s server processors and chipsets into the mainframe league.
Yes, you read that right. I said mainframe -- out loud. Hear me out; Hester’s really on to something. Half of his plan benefits clients and servers alike: AMD will grow technologies in areas where it already leads, like fast maths, multiple cores per chip, and extraordinarily fast bus architectures. But Hester detailed plans to extend AMD64 beyond the bounds of x86 to go after mainframe-grade “-ilities” like reliability, scalability, manageability, availability, and serviceability. AMD is also targeting hardware support for fine-grained resource allocation and utilisation through partitioning, which is the arbitrary assignment of a system’s physical assets like CPUs and memory to given tasks (physical subsystems within a physical system), and virtualisation, which allows each physical partition to host multiple systems virtualised in software.
x86 servers can support subsets of these features today, but no two approaches are the same, and each has gaps and limitations. The AMD64 server architecture will bake the “-ilities,” along with partitioning, virtualisation, and other big iron goodness, into AMD64 CPUs and chipsets. Imagine knowing that if you bought an Opteron II (or whatever) server, of any brand, you’d have a box that, if you chose to throw the switch on it, could turn itself into a rack-mountable mainframe that runs AMD64 editions of Linux, Windows, Solaris, or BSD, or all of them together. They won’t have to be architecture-aware to function well -- that’s the benefit of a standardised platform -- but the more aware the software is, the more advantage it can take of the environment.
Can AMD do it? Phil Hester hails from IBM, from whence bullet-proof things are known to come, and, more recently, from Newisys, the nearly invisible but inestimably influential think tank/engineering shop that’s already got a great handle on the whole “honey, I shrunk the mainframe” thing. Not coincidentally, the best of Hester’s IBM high-brightness brain trust landed at Newisys. Also not a coincidence is AMD’s cross-licensing, shared R&D pact with IBM on the Buck Rogers semiconductor science required to pull off the kind of long shot AMD has in mind.
I could be indulging in sheer hallucinatory fantasy (my speciality) to imagine that in four years or so (AMD isn’t standing still, but it’s also not rushed), we’ll be able to choose between compact, power-efficient, high performance, exquisitely interconnected rack-mount boxes and 64-way monolithic servers that you set up and manage exactly like the smaller rack units, and that run the same OSes and applications as the littler guys.
Now, this may sound to you like technology overkill, optimistic speculation, or maybe a big scary thunderhead of change. But trust me on this: you want x86 servers designed on standardised mainframe blueprints.